Commodity sales prediction method and device based on multi-modal feature fusion, and storage medium

Abstract

The invention relates to a commodity sales volume prediction method and device based on multi-modal feature fusion and a storage medium, and is applied to the technical field of store commodity sales volume prediction, and the method comprises the steps: obtaining static and dynamic multi-modal features affecting the store commodity sales volume, and carrying out the unique coding representation of a store and a commodity; the feature processing layer converts unique coding representation into dynamic embedded features and then performs weighted fusion on the dynamic embedded features and static features through feature gating, the feature gating has the advantage of dynamic weight distribution, and adaptive adjustment cannot be performed according to input sample features due to the fact that static weight distribution is adopted in a traditional method; then, the dynamic features continue to be fused through a cross attention mechanism, and soft alignment and weighted fusion of cross-modal features are achieved through the attention mechanism; then, the incidence relation between fusion features is learned through a multi-layer feedforward neural network of a calculation layer; and inputting the final fusion feature into a task layer to obtain a sales prediction result.

Description

Technical Field

[0001] This invention relates to the field of store merchandise sales forecasting technology, specifically to a merchandise sales forecasting method, apparatus, and storage medium based on multimodal feature fusion. Background Technology

[0002] In today's highly competitive market, sales forecasting is crucial for stores. By predicting future sales, stores can more scientifically plan replenishment, reducing capital tied up and losses due to inventory buildup, avoiding stockouts that impact sales, and improving inventory turnover. It also ensures a sufficient supply of popular items, reducing customer churn caused by stockouts and improving the shopping experience. Furthermore, it provides data for optimizing product mix, adjusting pricing strategies, and planning promotional activities, making operational decisions more aligned with actual needs. By reducing unnecessary inventory and logistical pressures, it optimizes the allocation of human resources and directly lowers operating costs. It helps stores identify market fluctuations (such as seasonal changes and unforeseen events) in advance, enabling them to develop contingency plans and stabilize operations. Accurate sales forecasting also aids in cash flow and profit estimation, providing a reliable foundation for store financial planning.

[0003] Traditional forecasting methods primarily employ time series analysis (such as ARIMA and exponential smoothing) or basic machine learning models (such as random forests and XGBoost), but they struggle to handle multi-source, heterogeneous data. With the rise of deep learning, models such as LSTM and Transformer have been introduced, but they suffer from issues like coarse feature fusion and rigid dynamic sequence processing, resulting in inaccurate sales forecasts. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, apparatus and storage medium for predicting commodity sales based on multimodal feature fusion, in order to solve the problems of existing technologies being unable to handle multi-source heterogeneous data, as well as coarse feature fusion and rigid dynamic sequence processing, which lead to inaccurate sales prediction results.

[0005] According to a first aspect of the present invention, a method for predicting product sales based on multimodal feature fusion is provided, the method comprising: Obtain static and dynamic features that affect the sales volume of products in the store; One-hot vector encoding is performed on the store and product to be predicted respectively. The one-hot vector encoding representations of the store and product are concatenated to obtain a unique fingerprint feature. The unique fingerprint feature, static feature, and dynamic feature are input into the feature processing layer of a pre-trained neural network model. The feature processing layer transforms the unique fingerprint feature into an embedding vector, which is used to convert sparse features into a set of dense features. The transformed dense features are then weighted and fused with the static features using a feature-gated feature fusion method to obtain the first fused feature. The feature processing layer further uses a cross-attention feature fusion method to fuse the first fused feature with the dynamic feature across modalities to obtain the second fused feature; The second fused feature is input into the computation layer of a pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector. Then, a multi-layer feedforward neural network is used to perform feature calculation on the fixed-length feature vector to obtain a third fused feature vector. The third fused feature vector is input into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction results of the store's products.

[0006] Preferably, The static features include product attributes and store attributes, which are constant features that do not change over time.

[0007] Preferably, The dynamic features include: external dynamic features, internal dynamic features, and time features; The external dynamic characteristics include: macroeconomic characteristics, holiday characteristics, online search index characteristics, and traffic characteristics within a preset range around the store; The internal dynamic features are sliding window features formed by historical data within the store, including the sales statistics of the current product in the current store over the past n weeks, the inventory quantity of the current product in the current store, the sales quantity of the product in the category of the current store, the sales quantity statistics, and the sales quantity and sales quantity statistics of the same product in other stores in the same area of ​​the store. The time feature refers to the week number of the current product's sales in the current store, used to indicate which stage of the product's life cycle it is currently in.

[0008] Preferably, The task layer adopts a composite pyramid architecture that combines primary and secondary task collaborative learning with feature sharing between tasks. The main task is the bottom layer of the composite pyramid, which is regression prediction of the sales volume of a single product in a single store. In addition to the main task at the bottom layer, each layer of the composite pyramid represents a secondary task. The tasks are combined together in a certain hierarchical manner. Each layer of the composite pyramid constructs a dedicated loss function to perform regression predictions for each auxiliary task and the main task.

[0009] Preferably, The auxiliary task includes a first branch task and a second branch task, and the last layer of the first branch task and the second branch task are respectively connected to the bottom layer where the main task is located. The third fusion feature vector is input into the top layer of the first branch task and the second branch task, respectively.

[0010] Preferably, The first branch task includes: First auxiliary task: Is there any sale of the current product in the current store at the current time? Second auxiliary task: The sales trend of the current products in the store compared to last week; Eighth auxiliary task: Which stage of the current product lifecycle does the current time belong to in the current store? The eighth auxiliary task is the superior task of the first and second auxiliary tasks, and the layer where the eighth auxiliary task is located receives the third fusion feature vector. The layers containing the first and second auxiliary tasks are respectively connected to the bottom layer containing the main task.

[0011] Preferably, The second side quest includes: Third auxiliary task: Preset the total sales volume of all products in all stores within the region for one week; Fourth auxiliary task: Total sales volume of all products in the current store for one week; Fifth auxiliary task: Preset the total sales volume of all stores within the region for the current product category over one week; Sixth auxiliary task: Total sales volume of products in the current category of the current store over the past week; Seventh auxiliary task: Preset the total sales volume of all products in all stores within the region for one week; The third auxiliary task is the superior task of the fourth and fifth auxiliary tasks, the superior task of the sixth auxiliary task is the fourth and fifth auxiliary tasks, and the superior task of the seventh auxiliary task is the fifth auxiliary task. The layer containing the third auxiliary task receives the third fused feature vector; The layers containing the sixth and seventh auxiliary tasks are respectively connected to the bottom layer containing the main task.

[0012] According to a second aspect of the present invention, a product sales prediction apparatus based on multimodal feature fusion is provided, the apparatus comprising: Feature acquisition module: used to acquire static and dynamic features that affect the sales volume of products in the store; Unique Encoding Representation Module: Used to perform one-hot vector encoding representations of the stores and products to be predicted, and to concatenate the one-hot vector encoding representations of the stores and products to obtain unique fingerprint features; The first feature fusion module is used to input the unique fingerprint feature, static feature, and dynamic feature into the feature processing layer of the pre-trained neural network model. The feature processing layer performs embedding vector transformation on the unique fingerprint feature to convert sparse features into a set of dense features. The transformed dense features are then weighted and fused with the static features using a feature-gated feature fusion method to obtain the first fused feature. The second feature fusion module is used in the feature processing layer to perform cross-modal feature fusion of the first fused feature and the dynamic feature through a cross-attention feature fusion method to obtain the second fused feature. The third feature fusion module is used to input the second fused feature into the computation layer of the pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector; then, it performs feature calculation on the fixed-length feature vector through a multi-layer feedforward neural network to obtain the third fused feature vector. Task prediction module: used to input the third fused feature vector into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction results of the store's products.

[0013] According to a third aspect of the present invention, a storage medium is provided, the storage medium storing a computer program, which, when executed by a host controller, implements the steps of the above-described method.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This application obtains multimodal features, including static and dynamic features, that influence the sales volume of store products and assigns unique codes to stores and products. These multimodal features and unique codes are then input into a pre-trained neural network model, which includes a feature processing layer, a computation layer, and a task layer. The feature processing layer first converts the unique codes into dynamically embedded features, and then uses feature gating to weightedly fuse the dynamically embedded features with static features. Feature gating offers the advantage of dynamic weight allocation, while traditional methods use static weight allocation, which cannot adaptively adjust based on the input sample features. Next, a cross-attention mechanism is used to further fuse dynamic features, achieving soft alignment and weighted fusion of cross-modal features. Then, a multi-layer feedforward neural network in the computation layer learns the correlation between the fused features. Finally, the fused features are input into the task layer to obtain the sales prediction result. This overcomes the problems of existing technologies, such as difficulty in handling multi-source heterogeneous data, coarse feature fusion, and rigid dynamic sequence processing.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] Figure 1 This is a flowchart illustrating a product sales forecasting method based on multimodal feature fusion according to an exemplary embodiment. Figure 2 This is a schematic diagram of primary and secondary task collaboration at the task layer, according to another exemplary embodiment. Figure 3 This is a system schematic diagram of a product sales forecasting device based on multimodal feature fusion, according to another exemplary embodiment. In the attached diagram: 1-Feature acquisition module, 2-Unique encoding representation module, 3-First feature fusion module, 4-Second feature fusion module, 5-Third feature fusion module, 6-Task prediction module. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0019] Example 1 Figure 1 This is a flowchart illustrating a multimodal feature fusion-based product sales prediction method according to an exemplary embodiment, such as... Figure 1 As shown, the method includes: S1, obtain the static and dynamic features that affect the sales volume of products in the store; S2, perform one-hot vector encoding representations for the store and product to be predicted respectively, and concatenate the one-hot vector encoding representations of the store and product to obtain a unique fingerprint feature; S3, the unique fingerprint feature, static feature and dynamic feature are input into the feature processing layer of the pre-trained neural network model. The feature processing layer performs embedding vector transformation on the unique fingerprint feature to convert sparse features into a set of dense features. The transformed dense features and static features are weighted and fused through a feature fusion method with feature gating to obtain the first fused feature. S4, the feature processing layer further performs cross-modal feature fusion of the first fused feature and the dynamic feature through a cross-attention feature fusion method to obtain the second fused feature; S5, the second fused feature is input into the computation layer of the pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector. Then, a multi-layer feedforward neural network is used to perform feature calculation on the fixed-length feature vector to obtain the third fused feature vector. S6, the third fused feature vector is input into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction result of the store's products; Understandably, the first step is to extract internal and external factors that potentially influence sales at various levels to construct features, including static and dynamic features, among which: Static features include product attributes (such as color, category, pricing, etc.) and store attributes (such as area, region, shelf type, etc.) that do not change over time; Dynamic features include external dynamic features, internal dynamic features, and time features; External dynamic characteristics include macroeconomic characteristics, holiday characteristics (the number of days between the current date and the Spring Festival, Double Eleven, etc.), online search index characteristics, and traffic characteristics around the store; The main macroeconomic characteristics include the PMI (Purchasing Managers' Index) and the Yiwu Small Commodities Index. The PMI is a comprehensive index compiled through monthly surveys of store purchasing managers, and it is calculated by weighting five sub-indices: new orders (30%), production (25%), employment (20%), supplier delivery time (15%), and raw material inventory (10%). The Yiwu Small Commodities Index consists of a price index, a prosperity index, and monitoring indicator indices, containing 23 sub-indices. The index adopts a multi-layered two-way weighted composite calculation method, and achieves integrated data collection, verification, and management through a distributed system. It covers dimensions such as online and offline transaction volume, price trends, merchant confidence, customer popularity, logistics services, and capital activity, forming a comprehensive indicator system for trade dynamic monitoring. Internal dynamic features are sliding window features formed from internal historical data, including the sales statistics of the current product in the current store over the past n weeks (including maximum, minimum, average, variance, etc.), the inventory quantity of the current product in the current store, the sales quantity of the product in the category in the current store, and the sales quantity statistics (describing the sales level of the product in the same category in the store), as well as the sales quantity and sales quantity statistics of the product in all stores in the region (describing the sales capacity of the product in the region of the store). The time feature indicates which week the product is currently being sold in the store, representing the current stage of the product's lifecycle. The product is represented by a one-hot vector code, and the store is represented by a one-hot vector code. After concatenation, a unique vector representation of the store and product combination is formed, which is referred to as the store and product fingerprint feature in this embodiment. Neural Network Section (Expert Network Design) The neural network consists of three main layers: the feature processing layer, the computation layer, and the task layer. Feature processing layer: First, the one-hot vector encoding of the fingerprint features is transformed by embedding, turning the sparse features into a set of dense features (dynamically embedded features). This set of dense features can uniquely represent a combination of stores and products. In the subsequent model training and learning process, this embedding vector can continuously learn the sales attributes of the combination of stores and products, which can be used to better describe the sales characteristics of the combination of stores and products. Static features and the aforementioned dense features of product and store combinations both describe the attributes of store products. Therefore, these two sets of features are first fused. During fusion, to avoid feature redundancy caused by direct splicing and thus leading to overfitting, this embodiment employs a FeatureGate feature fusion method, the mathematical expression of which is: Output=(E⊙σ(Wg·StaticFeat)+StaticFeat⊙(1-σ(Wg·StaticFeat)) In the formula, E represents the dynamic embedding feature, StaticFeat is the static feature constructed above, σ is the sigmoid activation function, Wg is the trainable weight matrix, and sigmoid is a gate unit that dynamically generates weight coefficients in the range of 0-1 to perform weighted filtering on the dynamic embedding feature, suppressing redundant information; and retaining the original static feature information to ensure that the basic features are not lost. Compared to traditional feature concatenation or fully connected layer fusion, the FeatureGating layer used in this embodiment has the advantage of dynamic weight allocation: traditional methods use static weight allocation, which cannot adaptively adjust according to the features of the input samples; this layer achieves sample-level dynamic weighting through sigmoid gating, improving the accuracy of feature fusion, while filtering low-information features through the gating mechanism to avoid overfitting caused by "feature stacking". Experiments show that it can reduce feature redundancy by 20%-30%.

[0020] The FeatureGate layer outputs a first fused feature that combines static features and embedding features; The first fused feature and the dynamic feature are input into the CrossAttention layer to achieve deep interaction of spatiotemporal features. The calculation formula is as follows: CrossAttention(Q,K,V)=softmax(QKt / dk)V In the formula, Q is the dynamic feature, K and V are the first fusion features constructed above, and dk is the dimension of each attention head, which is preset to 128. This design enables deep interaction between dynamic and static features, and dynamic learning through the attention mechanism to determine which static features should be given more attention under different time features. The CrossAttention layer outputs a second fused feature that combines dynamic features with the first fused feature; Computation layer: The computational layer employs temporal attention pooling to compress the second fused feature into a fixed-length feature vector. This vector is then combined with a multi-layer feedforward neural network to perform feature computation, which is used to learn the correlation between features, resulting in a set of fixed-length feature vectors, i.e., the third fused feature. Task layer: In constructing the loss function, this embodiment designs a composite pyramid architecture of main-auxiliary task collaborative learning and inter-task feature sharing. Each layer constructs a separate loss function, and each layer contains an auxiliary task. The bottom layer of the pyramid is the main task, namely, regression prediction of the sales quantity of a single product in a single store; as shown in the appendix. Figure 2 As shown, the auxiliary tasks include whether the store's products should have sales at the current time (binary classification), which stage of a product's lifecycle it is in (multi-class classification), the sales trend compared to the previous time (multi-class classification), and hierarchical sales prediction, covering multiple levels of sales regression prediction such as category level and global level. Tasks are combined in a certain hierarchical manner, and knowledge transfer is achieved through feature contribution. For example, the global sales prediction task is the superior task of the store-level sales prediction task, and the subordinate tasks acquire knowledge of the superior task by sharing the features generated by the superior task. The superior task provides directional constraints for the subordinate tasks. As attached Figure 2 As shown, the auxiliary task comprises two branches. One branch consists of auxiliary tasks 1, 2, and 8, where auxiliary task 8 is the superior task of auxiliary tasks 1 and 2. The other branch consists of auxiliary tasks 3, 4, 5, 6, and 7, where auxiliary task 3 is the superior task of auxiliary tasks 4 and 5, auxiliary task 6 is the superior task of auxiliary tasks 4 and 5, and auxiliary task 7 is the superior task of auxiliary task 5. Auxiliary tasks 3 and 8 form the top layer of the composite pyramid and receive the third fusion feature mentioned above. Auxiliary tasks 1, 2, 6, and 7 are the superior tasks of the main task, meaning the output of the layer containing auxiliary tasks 1, 2, 6, and 7 serves as the input to the final main task. The main task layer outputs the predicted sales quantity data for the current product in the current store.

[0021] Example 2 Figure 3 This is a system schematic diagram of a multimodal feature fusion-based product sales forecasting apparatus according to another exemplary embodiment, the apparatus comprising: Feature acquisition module 1: Used to acquire static and dynamic features that affect the sales volume of products in the store; Unique Encoding Representation Module 2: Used to perform one-hot vector encoding representations of the stores and products to be predicted, and to concatenate the one-hot vector encoding representations of the stores and products to obtain unique fingerprint features; First feature fusion module 3: used to input the unique fingerprint feature, static feature and dynamic feature into the feature processing layer of the pre-trained neural network model. The feature processing layer transforms the unique fingerprint feature into an embedding vector, which is used to convert sparse features into a set of dense features. The transformed dense features and static features are weighted and fused through a feature gating feature fusion method to obtain the first fused feature. Second feature fusion module 4: The feature processing layer further uses a cross-attention feature fusion method to perform cross-modal feature fusion of the first fused feature and the dynamic feature to obtain the second fused feature; The third feature fusion module 5 is used to input the second fused feature into the computation layer of the pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector; then, it performs feature calculation on the fixed-length feature vector through a multi-layer feedforward neural network to obtain the third fused feature vector. Task prediction module 6: used to input the third fused feature vector into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction result of the store's products.

[0022] Example 3 This embodiment provides a storage medium storing a computer program, which, when executed by a host controller, implements the various steps in the above method. It is understood that the storage medium mentioned above can be a read-only memory, a hard disk, or an optical disk, etc.

[0023] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0024] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0025] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0026] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0027] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0028] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0029] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0030] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0031] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A product sales prediction method based on multimodal feature fusion, characterized in that, The method includes: Obtain static and dynamic features that affect the sales volume of products in the store; One-hot vector encoding is performed on the store and product to be predicted respectively. The one-hot vector encoding representations of the store and product are concatenated to obtain a unique fingerprint feature. The unique fingerprint feature, static feature, and dynamic feature are input into the feature processing layer of a pre-trained neural network model. The feature processing layer transforms the unique fingerprint feature into an embedding vector, which is used to convert sparse features into a set of dense features. The transformed dense features are then weighted and fused with the static features using a feature-gated feature fusion method to obtain the first fused feature. The feature processing layer further uses a cross-attention feature fusion method to fuse the first fused feature with the dynamic feature across modalities to obtain the second fused feature; The second fused feature is input into the computation layer of a pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector. Then, a multi-layer feedforward neural network is used to perform feature calculation on the fixed-length feature vector to obtain a third fused feature vector. The third fused feature vector is input into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction results of the store's products.

2. The method according to claim 1, characterized in that, The static features include product attributes and store attributes, which are constant features that do not change over time.

3. The method according to claim 2, characterized in that, The dynamic features include: external dynamic features, internal dynamic features, and time features; The external dynamic characteristics include: macroeconomic characteristics, holiday characteristics, online search index characteristics, and traffic characteristics within a preset range around the store; The internal dynamic features are sliding window features formed by historical data within the store, including the sales statistics of the current product in the current store over the past n weeks, the inventory quantity of the current product in the current store, the sales quantity of the product in the category of the current store, the sales quantity statistics, and the sales quantity and sales quantity statistics of the same product in other stores in the same area of ​​the store. The time feature refers to the week number of the current product's sales in the current store, used to indicate which stage of the product's life cycle it is currently in.

4. The method according to claim 3, characterized in that, The task layer adopts a composite pyramid architecture that combines primary and secondary task collaborative learning with feature sharing between tasks. The main task is the bottom layer of the composite pyramid, which is regression prediction of the sales volume of a single product in a single store. In addition to the main task at the bottom layer, each layer of the composite pyramid represents a secondary task. The tasks are combined together in a certain hierarchical manner. Each layer of the composite pyramid constructs a dedicated loss function to perform regression predictions for each auxiliary task and the main task.

5. The method according to claim 4, characterized in that, The auxiliary task includes a first branch task and a second branch task, and the last layer of the first branch task and the second branch task are respectively connected to the bottom layer where the main task is located. The third fusion feature vector is input into the top layer of the first branch task and the second branch task, respectively.

6. The method according to claim 5, characterized in that, The first branch task includes: First auxiliary task: Is there any sale of the current product in the current store at the current time? Second auxiliary task: The sales trend of the current products in the store compared to last week; Eighth auxiliary task: Which stage of the current product lifecycle does the current time belong to in the current store? The eighth auxiliary task is the superior task of the first and second auxiliary tasks, and the layer where the eighth auxiliary task is located receives the third fusion feature vector. The layers containing the first and second auxiliary tasks are respectively connected to the bottom layer containing the main task.

7. The method according to claim 6, characterized in that, The second side quest includes: Third auxiliary task: Preset the total sales volume of all products in all stores within the region for one week; Fourth auxiliary task: Total sales volume of all products in the current store for one week; Fifth auxiliary task: Preset the total sales volume of all stores within the region for the current product category over one week; Sixth auxiliary task: Total sales volume of products in the current category of the current store over the past week; Seventh auxiliary task: Preset the total sales volume of all products in all stores within the region for one week; The third auxiliary task is the superior task of the fourth and fifth auxiliary tasks, the superior task of the sixth auxiliary task is the fourth and fifth auxiliary tasks, and the superior task of the seventh auxiliary task is the fifth auxiliary task. The layer containing the third auxiliary task receives the third fused feature vector; The layers containing the sixth and seventh auxiliary tasks are respectively connected to the bottom layer containing the main task.

8. A product sales forecasting device based on multimodal feature fusion, characterized in that, The device includes: Feature acquisition module: used to acquire static and dynamic features that affect the sales volume of products in the store; Unique Encoding Representation Module: Used to perform one-hot vector encoding representations of the stores and products to be predicted, and to concatenate the one-hot vector encoding representations of the stores and products to obtain unique fingerprint features; The first feature fusion module is used to input the unique fingerprint feature, static feature, and dynamic feature into the feature processing layer of the pre-trained neural network model. The feature processing layer performs embedding vector transformation on the unique fingerprint feature to convert sparse features into a set of dense features. The transformed dense features are then weighted and fused with the static features using a feature-gated feature fusion method to obtain the first fused feature. The second feature fusion module is used in the feature processing layer to perform cross-modal feature fusion of the first fused feature and the dynamic feature through a cross-attention feature fusion method to obtain the second fused feature. The third feature fusion module is used to input the second fused feature into the computation layer of the pre-trained neural network model. The computation layer first uses temporal attention pooling to compress the second fused feature into a fixed-length feature vector; then, it performs feature calculation on the fixed-length feature vector through a multi-layer feedforward neural network to obtain the third fused feature vector. Task prediction module: used to input the third fused feature vector into the task layer of the pre-trained neural network model, and the task layer outputs the sales prediction results of the store's products.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by the main controller, implements each step of the multimodal feature fusion method for predicting commodity sales as described in any one of claims 1-7.

Images